JP5756443B2 - Image classification device, image identification device, and program - Google Patents

Description

  The present invention relates to an image classification device, an image identification device, and a program.

  Techniques for allowing a device such as a computer to identify patterns or patterns such as clothes have been studied (for example, Non-Patent Documents 1 to 3). Non-Patent Document 1 discloses a method based on template matching. This template matching method is suitable for identifying exactly the same pattern. However, in this method, when the width of the stripe pattern is different even in the category of the same stripe pattern, or when the size or interval of the polka dots is different in the category of the same polka dot pattern, it is difficult to distinguish these from the same category. In addition, in order to deal with various patterns, there is a drawback in that in generating a dictionary for storing patterns in advance, it is necessary to classify the patterns and collect a large number of images that are scaled and aligned for each type of pattern. is there. Furthermore, there is a drawback that it is also vulnerable to deformation and distortion of clothes. For this reason, in this method, clothing is divided into fine patch areas, and affine transformation is performed for each area.

Non-Patent Document 2 shows a method using frequency analysis and a method using local features. This frequency analysis method is considered effective in identifying a relatively simple geometric pattern, that is, a specific simple pattern such as a striped pattern. In addition, it is considered easy to distinguish the stripe pattern from the check. However, if there is a change in the resolution of the image, or if you want to identify both striped patterns with different intervals as the same category of striped patterns, if you want to identify them as the same category, it is difficult to identify them as the same category .
Further, the method using local features is suitable for identifying fine patterns. Also, it is strong against changes in scale. However, it is considered weak against pattern distortion and noise. Well, it is not suitable for extracting global features such as striped patterns.

  Non-Patent Document 3 discloses a method based on texture analysis. This texture analysis method can detect the directionality of a repetitive pattern in an image, and is suitable for identifying a fine repetitive pattern, for example, an image obtained by enlarging a cloth. However, it is not suitable for identifying complex patterns including various changes, such as floral patterns and animal patterns, rather than repeating patterns in fine units. In addition to the bit run analysis described in Non-Patent Document 3, the texture analysis includes a method using a co-occurrence matrix, a moment feature, and the like.

Tetsunori Taguchi, Teruyoshi Aoki, Hiroshi Yasuda, “A Study on Cloth Shape Recognition for“ MIRACLE ”System”, IEICE Technical Report IE2002-113, pp. 55-58, 2002 Junpei Yamaguchi, Kazutaka Shimada, Shuichi Hamada, Toshiro Ejima, Tsutomu Endo, “Person Identification Based on Facial Features and Context Information”, Journal of Japan Society for Fuzzy Intelligence, Vol. 23, no. 2, pp. 137-145, 2011 Toshio Mori, Kazumi Ide, “Texture characteristics and bit-run analysis of lace patterns”, Bulletin of Gifu Women's University, No. 30, pp. 161-166, 2010

  However, in the methods shown in Non-Patent Documents 1 to 3, in order to stably create a dictionary storing pattern features and the like and to obtain identification results and the like, data for creating the dictionary and identification areas are cut out. It is necessary to select the part without the crease of the clothes and the boundary between the outline and the background at the time, but it is troublesome to do this manually, and it is difficult to prepare a large amount of data .

  The present invention has been made to solve the above problems, and an object of the present invention is to provide an image classification apparatus capable of stably obtaining a dictionary creation and identification result and improving pattern identification accuracy. And an image identification apparatus and a program.

  In order to solve the above problem, the image classification device according to the present invention is an image classification device in which a learning image and a plurality of combinations of labels indicating patterns included in the learning image are input as learning data, the learning image First, a region where the fabric pattern imaged is discontinuous or a region where the fabric pattern is wrinkled is detected, and the region other than the region in the learning image and including the fabric pattern is cut out. An image cutout unit, a first feature quantity extraction unit that extracts a plurality of feature quantities indicating features of a pattern included in an image cut out by the first image cutout unit, and the first feature quantity extraction unit extracted A representative vector extracting unit that extracts a representative vector for each of a plurality of feature amounts; and a histogram of feature amounts corresponding to the representative vector based on the representative vector extracted by the representative vector extracting unit. An identification parameter that is a weighting coefficient for the histogram based on the histogram generated by the feature histogram generation unit, the histogram generated by the first feature histogram generation unit, and the label corresponding to the image, And a learning unit that generates pattern classification data in which representative vectors corresponding to the identification parameters are associated.

  The image classification device according to the present invention is an image classification device in which a plurality of combinations of a learning image and a label indicating a pattern included in the learning image are input as learning data, and the fabric imaged in the learning image A first image cutout unit that detects a discontinuous region or a region having a fabric wrinkle and cuts out a region other than the region in the learning image and including the fabric pattern; A first feature amount extraction unit that extracts a plurality of feature amounts indicating features of a pattern included in an image cut out by the first image cutout unit and a plurality of feature amounts extracted by the first feature amount extraction unit are connected. A representative vector extracting unit that extracts a representative vector for each feature amount obtained in this manner, and generating a histogram of feature amounts corresponding to the representative vector based on the representative vector extracted by the representative vector extracting unit. Based on a first feature histogram generation unit, a histogram generated by the first feature histogram generation unit, and a label corresponding to the image, an identification parameter that is a weighting factor for the histogram is calculated, and the calculated identification A learning unit that generates pattern classification data in which a parameter and a representative vector corresponding to the identification parameter are associated with each other is provided.

  Further, the image identification device according to the present invention detects an area where the pattern of the fabric imaged in the input image is discontinuous or an area where the cloth is wrinkled, and is an area other than the area in the image. A second image cutout unit that cuts out an area including the pattern of the fabric, and a second feature amount extraction unit that extracts a plurality of feature amounts indicating features of the pattern included in the image cut out by the second image cutout unit And a second feature for generating a histogram for each of the plurality of feature amounts extracted by the second feature amount extraction unit based on the representative vector included in the pattern classification data generated by the image classification device described above. Based on the histogram generation unit, the identification parameter included in the pattern classification data stored in the image classification device, and the histogram generated by the second feature histogram generation unit, the input Characterized in that it comprises an identification portion for identifying the pattern of the fabric of the image.

The program in the present invention is a program for causing a computer to function as the image classification device described above.
The program in the present invention is a program for causing a computer to function as the image identification device described above.

  According to the present invention, a region with little distortion such as a discontinuous pattern or a wrinkled region is cut out from an image, a plurality of feature amounts are extracted from the cut out image, and the extracted feature amounts are combined and identified. By calculating the parameters, dictionary creation and identification results can be obtained stably. In addition, by identifying a pattern using an identification parameter obtained based on an image of a region with less distortion, the accuracy of identifying the pattern can be improved.

It is a schematic block diagram which shows the structure of the image classification apparatus 1 in this embodiment. 4 is a diagram illustrating an example of processing performed by an image cutout unit 11 according to the embodiment using Method Example 1. FIG. It is a schematic block diagram which shows the structural example of the image identification device 2 in the embodiment. It is a figure which shows an example of the identification result of the pattern by the image identification apparatus 2 in the embodiment. It is a schematic block diagram which shows the structure of 1 A of image classification apparatuses as a modification of the image classification apparatus 1. FIG.

Hereinafter, an image classification device, an image identification device, and a program according to an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic block diagram showing a configuration of an image classification device 1 in the present embodiment. The image classification device 1 inputs, as learning data, an image (learning image) obtained by capturing clothes, bags, and the like, and a label that identifies the pattern (or pattern) of the fabric such as clothes and bags captured in the image. Then, pattern classification data which is information for identifying the pattern is generated and stored. As shown in the figure, the image classification apparatus 1 according to the present embodiment includes an image cutout unit 11, a feature extraction unit 12, a representative vector extraction unit 13, a feature histogram generation unit 14, a learning unit 15, and a representative vector / identification parameter. A storage unit 16 is provided.

  The image cutout unit 11 selects and cuts out an area from which the texture can be stably obtained without being affected by the shape or wrinkles of clothes or the like included in the input image. The image cutout unit 11 outputs the cutout image to the feature extraction unit 12. In other words, the image cutout unit 11 detects a region where the fabric pattern in the image is discontinuous or a region where the fabric pattern is wrinkled, and is a region other than the region in the image and including the fabric pattern. Cut out the image. In the following, three examples of methods in which the image cutout unit 11 cuts out an area with less wrinkles from an input image are shown.

(Method Example 1)
In the method shown in Method Example 1, a region excluding the influence of a shape such as a sleeve and a neck is cut out using a morphological process. Specifically, first, the input image is binarized. An opening of a morphological process is performed on the binarized image. The opening in the morphological process is a process of removing uneven noise by performing dilation after erosion.
Using the binary image obtained by the opening as a mask, a region with little wrinkles is cut out from the input image, and the cut out image is output.

FIG. 2 is a diagram illustrating an example of processing performed by the image cutout unit 11 according to the present embodiment using the method example 1. In the same figure, image example 1 shows an example in which a region with little wrinkles is cut out from a striped cut-and-sew image (A). An image (B) is generated by binarization of the image (A), and an opening is performed on the image (B) to obtain an image (C). By using this image (C) as a mask, an image (D) obtained by cutting out a region with little wrinkles from the image (A) can be obtained.
Further, image example 2 shows an example in which an area with less wrinkles is cut out from an image (E) of a checkered shirt. Similarly to the image example 1, the image (F) is generated by binarizing the image (E), and the image (F) is opened to obtain the image (G). An image (H) obtained by cutting out an area with few wrinkles from the input image (E) using the image (G) as a mask can be obtained.

(Method Example 2)
In Method Example 2, mainly in the case of clothing, the collar, shape, hem, etc. of clothing are roughly aligned for each category of shape, such as jacket or dress, using a probability model fitting of clothing images by a known technique. be able to. Such techniques are described in, for example, Reference 1 (Reiko Sudo, Jun Shimamura, Masashi Morimoto, “Proposal of Shape Matching Method Using Procrustes Analysis and Application to Similarity Determination in Clothing Images”, Science Technique, vol. 11, No. 353, PRMU 2011-139, pp. 83-88, 2011). As a result of the alignment, it is possible to estimate which feature point of the clothing is based on which distribution of the probability model the feature point of the clothing image belongs to. Therefore, by storing in advance information indicating the distribution of the area with less wrinkles for each category, the feature points of the area with less wrinkles are specified, and the surrounding area including the specified feature points is cut out from the image. As a result, it is possible to obtain an image obtained by cutting out a region with less wrinkles from the input image.

(Method Example 3)
A region having few wrinkle features is cut out from the input image by using a known technique for detecting wrinkle features. For this, a frequency component when a wrinkle is generated based on a similar sample image is learned in advance, and a technique for extracting a region with a small frequency component or detecting a wrinkle feature using a known technique is used. Thus, it is possible to extract a region having a small wrinkle feature. Known techniques include, for example, Reference 2 (Kintoshi Yamazaki, Masayuki Inaba, “Clothing Discovery Method Based on Wrinkle Features for Life Support Robots”, Proceedings of the Annual Conference of the Robotics Society of Japan, 2R1-05, 2009) There are techniques described in.

The feature extraction unit 12 includes k (k ≧ 2) feature extraction units that extract different features from the image cut out by the image cutout unit 11. In the example illustrated in FIG. 1, the feature extraction unit 12 includes a texture feature extraction unit 121, a color feature extraction unit 122, a shape feature extraction unit 123, and the like.
The texture feature extraction unit 121, the color feature extraction unit 122, the shape feature extraction unit 123, and the like extract feature quantities indicating features related to appearance attributes such as clothes. For example, the texture feature extraction unit 121 extracts local feature amounts such as SIFT (Scale Invariant Feature Transform), Hog feature amounts, gradient features, frequency features, and the like from the image with respect to the fabric pattern captured in the image.

The color feature extraction unit 122 outputs a histogram of an image cut out by the image cut-out unit 11 in a color system (color model) such as RGB, Lab, or HSV with respect to a fabric pattern captured in the image.
The shape feature extraction unit 123 extracts a feature amount that represents a feature of the outline shape of the entire clothing. For example, the shape feature extraction unit 123 extracts a distance between a template of a specific contour shape prepared in advance and a shape included in the image, a fractal dimension indicating complexity, a feature amount based on curvature, and the like.

  By using a combination of feature quantities obtained in accordance with different features in this way, there is a possibility that the accuracy of pattern recognition may be improved when there is a high correlation between the shape of the overall clothing and the pattern.

  The representative vector extraction unit 13 includes k vector extraction units 131 to 13k corresponding to the k feature amounts extracted from the image by the feature extraction unit 12. For example, in the example shown in FIG. 1, the feature amount extracted by the texture feature extraction unit 121 is input to the vector extraction unit 131, and the feature amount extracted by the color feature extraction unit 122 is input to the vector extraction unit 132. The feature amount extracted by the shape feature extraction unit 123 is input to the vector extraction unit 133.

  Each of the vector extraction units 131 to 13k clusters each feature amount obtained in the feature extraction unit 12 and extracts a representative vector. For example, when the feature quantity extracted by the texture feature extraction unit 121 is a SIFT feature quantity, the representative vector extracted by the vector extraction unit 131 corresponds to the Visual Word of the Bag-of-Feature method. Here, the representative vectors are obtained by clustering the feature amounts by the vector extraction units 131 to 13k for each feature amount obtained in each unit of the feature extraction unit 12.

The feature histogram generation unit 14 includes k histogram generation units 141 to 14k corresponding to the representative vectors extracted by the representative vector extraction unit 13. The representative vector extracted by the vector extraction unit 13i is input to the histogram generation unit 14i (i = 1, 2,..., K). The histogram generation unit 14i (i = 1, 2,..., K) generates a histogram having the input representative vector as the bin axis.
The concept of processing in the representative vector extraction unit 13 and the feature histogram generation unit 14 described above corresponds to the Bag-of-Feature method. Regarding the Bag-of-Feature technique, for example, Reference 3 (Tomoyuki Nagahashi, Yuji Ihara, Hironobu Fujiyoshi, “Image Classification by Bag-of-feature Using Co-occurrence Expression of Foreground and Background Information”, Recognition / understanding symposium MIRU 2010, July 2010).

The learning unit 15 learns an identification parameter for classifying a pattern category such as clothes while referring to a label input as learning data together with an image. As a classifier used in the learning unit 15, for example, SVM (Support Vector Machine) can be used. In this case, assuming that there are six patterns of patterns A to F, each of six types, such as an identification parameter indicating the pattern of A, an identification parameter indicating the pattern of B,. The identification parameter is learned. In other words, the learning unit 15 calculates an identification parameter, which is a weighting coefficient corresponding to the label, based on the input label and the histogram of each feature amount generated by the feature histogram generation unit 14. The learning unit 15 generates pattern classification data obtained by learning using a large number of different learning data, in which the identification parameters are associated with the representative vectors, and the generated pattern classification data is used as the representative vector / identification parameter storage unit 16. Remember me.
The representative vector / identification parameter storage unit 16 stores pattern classification data in which the distance between feature histograms, which is the distance between feature histograms generated by the feature histogram generation unit 14, is associated with the identification parameter and the representative vector. Also good.

  FIG. 3 is a schematic block diagram illustrating a configuration example of the image identification device 2 in the present embodiment. The image identification device 2 identifies the pattern included in the input image based on the pattern classification data generated by the image classification device 1, that is, the pattern classification data stored in the representative vector / identification parameter storage unit 16. Information indicating the category corresponding to the pattern of the input image is output. As shown in FIG. 3, the image identification device 2 includes an image cutout unit 21, a feature extraction unit 22, a feature histogram generation unit 23, and an identification unit 24.

  Similar to the image cutout unit 11 provided in the image classification device 1, the image cutout unit 21 selects a region in the input image that can stably obtain a texture without being affected by the shape or wrinkles of clothes or the like. To do. The image cutout unit 21 cuts out the selected region and outputs an image of the cutout region to the feature extraction unit 22.

  Similar to the feature extraction unit 12, the feature extraction unit 22 includes a feature extraction unit that extracts the feature for each of k different features. In the example shown in FIG. 3, the feature extraction unit 22 includes a texture feature extraction unit 221, a color feature extraction unit 222, a shape feature extraction unit 223, and the like. The texture feature extraction unit 221, the color feature extraction unit 222, and the shape feature extraction unit 223 have the same configuration as the texture feature extraction unit 121, the color feature extraction unit 122, and the shape feature extraction unit 123 shown in FIG. have.

  Similar to the feature histogram generation unit 14, the feature histogram generation unit 23 includes k histogram generation units 231 to 23 k corresponding to the feature amounts extracted by the feature extraction unit 22. The histogram generation unit 23i (i = 1, 2,..., K) is a representative vector and an identification parameter corresponding to the input feature quantity, and is stored as pattern classification data in the representative vector / identification parameter storage unit 16. Read representative vectors and identification parameters. The histogram generation unit 23 i generates a histogram having the read representative vector as the axis of the bin from the feature amount input from the feature extraction unit 22. The histogram generation unit 23 i outputs the generated histogram of the feature amount to the identification unit 24.

  The identification unit 24 includes m classifiers 241 to 24m corresponding to the number (m) of patterns (patterns) included in the pattern classification data stored in the representative vector / identification parameter storage unit 16. The classifiers 241 to 24 m are provided for each identification parameter included in the pattern classification data stored in the representative vector / identification parameter storage unit 16, in other words, for each pattern to be identified by the image identification device 2. Each of the classifiers 241 to 24 m adds itself to the input image based on the pattern classification data stored in the representative vector / discrimination parameter storage unit 16 and the histogram of each feature amount generated by the feature histogram generation unit 23. It is determined whether a pattern corresponding to is included, and the determination result is output. Each discriminator 241 to 24m uses the same discrimination method as that used in the learning unit 15, and corresponds to the input image from the discrimination parameter included in the pattern classification data and the histogram of the feature amount. It is determined whether or not a pattern is included.

For example, if the classifiers 241 to 24m are SVM and there are six patterns of patterns A to F, the representative vector / identification parameter storage unit 16 stores the identification parameter of the A pattern and the identification parameter of the B pattern. ,..., F are stored as representative vectors corresponding to the identification parameters of the F pattern. That is, the image classification device 1 learns the patterns A to F, and a combination of the representative vector and the identification parameter for each pattern obtained by learning is stored in the representative vector / identification parameter storage unit 16. In this case, the classifiers 241 to 246 provided in the image identification device 2 correspond to patterns A to F. Each discriminator 241 to 246 outputs a predetermined value (for example, “1”) when a pattern corresponding to itself is included in the input image, and a different value when the pattern does not include the pattern. (For example, “0”) is output. The identification unit 24 identifies a pattern included in the input image based on the outputs of the classifiers 241 to 246, and outputs information indicating the pattern.
In addition, the output of each discriminator 241 to 24m is a binary value of 1 or 0, and constitutes a multi-class discriminator by a 2-class discriminator, and has a probability (continuous value) that each pattern is included in the input image. It is also possible. In that case, the identification unit 24 outputs the probability that each pattern is included in the input image based on the outputs of the classifiers 241 to 246. For example, it is a sequence of numbers “0.1, 0.05, 0.15, 0.6, 0.04, 0.06”, etc., which becomes 1 in total.

Note that “1” may be output from the plurality of discriminators 241 to 24m. In such a case, information indicating a plurality of patterns is output. In order to determine the pattern output by the image identification device 2 as one pattern, the image cutout unit 21 cuts out a plurality of different areas, and the output results of the classifiers 241 to 24m obtained based on the images of the plurality of areas. The majority vote may be taken.
Moreover, when making the output result of the discriminators 241-24m into a continuous value, let the pattern with the largest numerical value be an identification result. For example, when the output of the identification unit 24 is the sequence of numbers in the above-described example, it is determined that the pattern D.

  FIG. 4 is a diagram illustrating an example of a pattern identification result by the image identification device 2 according to the present embodiment. Note that the pattern classification data used by the image identification device 2 is a pattern label such as border (dotted pattern) or dots (polka dot) when the image classification device 1 learns, that is, when the image classification device 1 generates pattern classification data. , Check (lattice pattern), zebra (suma pattern), leopard (leopard pattern), python (snake pattern), and liberty (small floral pattern). The texture feature extraction units 121 and 221 extract feature amounts using SIFT, and the color feature extraction units 122 and 222 extract feature amounts using RGB. The shape feature extraction units 123 and 223 are not used. The image as an object for identifying the pattern is an image of clothes displayed on a web purchase site on the Internet.

  As shown in FIG. 4, it is difficult to identify the same pattern as the same pattern in the past. Different stripes and lattice patterns with different widths are identified as the same stripe pattern and lattice pattern. It can be confirmed that it can be recognized as. Moreover, it can be confirmed that complicated animal patterns and small floral patterns can be recognized considerably.

By using the image classification device 1 and the image identification device 2 in the present embodiment, it is possible to improve the accuracy of identifying a pattern of a product such as clothes or a bag.
Further, as shown in FIG. 4, a striped pattern can be identified regardless of the width and interval of the stripes, or a polka dot pattern can be identified regardless of the size and density of the polka dots. Since the pattern can be identified by absorbing variations within the category, the image classification device 1 and the image identification device 2 are suitable for classifying a product pattern for use such as search.
In addition, the image classification device 1 and the image identification device 2 include the feature extraction units 12 and 22 that extract a plurality of features, and the pattern is identified based on the extracted features. The pattern can be identified in consideration of the correlation between the pattern and the pattern and the pattern and the shape, and the pattern identification accuracy can be improved.

Further, since the image classification device 1 and the image identification device 2 include the image cutout units 11 and 21, even if the clothes in the input image have wrinkles or distortion, the wrinkle and distortion regions are excluded. Since an image can be used, pattern identification accuracy can be improved.
In addition, the feature extraction units 12 and 22 extract feature amounts corresponding to a plurality of features, for example, image scaling and resolution, which are disadvantages of the techniques described in Non-Patent Document 1 and Non-Patent Document 2, for example. It is possible to suppress a decrease in pattern identification accuracy caused by.

  Further, the image classification apparatus 1 includes a learning unit 15 that learns representative vectors based on feature quantities of a plurality of features including features other than patterns, for example, features of colors and shapes, and their weights. The identification accuracy can be improved as compared with the case of identifying only the handle alone. Also, by distinguishing using multiple features including features other than patterns, floral patterns are often found in red clothes, animal patterns are often found in brown clothes, and striped patterns are often found in cold clothes. Trends such as, floral patterns are often found in clothes with a skirt or one-piece shape, but not in clothes with a jacket shape, and check patterns are often found in shirt-shaped clothes but less in clothes with a trouser shape. If there is such a tendency, the pattern identification accuracy can be increased.

(Modification)
In the image classification device 1 illustrated in FIG. 1, the configuration in which the vector extraction units 131 to 13 k extract the representative vector for each feature amount extracted by the feature extraction unit 12 has been described. However, the present invention is not limited to this, and the representative vector may be extracted using a feature amount obtained by connecting the feature amounts extracted by the feature extraction unit 12.
FIG. 5 is a schematic block diagram showing a configuration of an image classification device 1A as a modification of the image classification device 1. In the image classification apparatus 1A shown in the figure, the difference from the image classification apparatus 1 is that the feature amounts extracted by the feature extraction unit 12 are combined and input to the vector extraction units 331 to 33k. . As described above, the combined feature amounts are concatenated and clustered as one feature amount in the vector extraction units 331 to 33k to obtain a representative vector. For example, when the texture feature amount is 128 dimensions and the color feature amount is 27 dimensions, the vector extraction units 331 to 33k perform clustering and calculation of representative vectors as 155-dimensional feature amounts obtained by combining these two feature amounts. . As a result, the pattern can be identified with higher accuracy. In this case, similarly in the image identification device 2, a histogram is generated based on the feature amount obtained by connecting the feature amounts extracted by the feature extraction unit 22.

  A program for realizing the functions of the image classification device and the image identification device according to the present invention is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read by a computer system and executed. Thus, the processing of each functional unit may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

DESCRIPTION OF SYMBOLS 1, 1A ... Image classification apparatus 2 ... Image identification apparatus 11, 21 ... Image clipping part 12, 22 ... Feature extraction part 13 ... Representative vector extraction part 14, 23 ... Feature histogram generation part 15 ... Learning part 16 ... Representative vector and identification Parameter storage unit 24 ... identification unit 33 ... representative vector extraction unit 121, 221 ... texture feature extraction unit 122, 222 ... color feature extraction unit 123, 223 ... shape feature extraction unit 131, 132, 133, 13k ... vector extraction unit 141, 142, 143, 14k, 231, 232, 23k ... Histogram generator 241, 242, 24m ... Discriminator 331, 332, 333, 33k ... Vector extractor

Claims (5)

An image classification device in which a plurality of combinations of learning images and labels indicating patterns included in the learning images are input as learning data,
An area in which the fabric pattern captured in the learning image is discontinuous or an area with fabric wrinkles is detected, and an area other than the area in the learning image that includes the fabric pattern is cut out. A first image cutout unit;
A first feature amount extraction unit that extracts a plurality of feature amounts indicating features of a pattern included in the image cut out by the first image cutout unit;
A representative vector extraction unit that extracts a representative vector for each of the plurality of feature amounts extracted by the first feature amount extraction unit;
A first feature histogram generation unit that generates a histogram of feature amounts corresponding to the representative vector based on the representative vector extracted by the representative vector extraction unit;
Based on the histogram generated by the first feature histogram generation unit and a label corresponding to the image, an identification parameter that is a weighting factor for the histogram is calculated, and the calculated identification parameter and a representative corresponding to the identification parameter An image classification apparatus comprising: a learning unit that generates pattern classification data associated with a vector. An image classification device in which a plurality of combinations of learning images and labels indicating patterns included in the learning images are input as learning data,
An area in which the fabric pattern captured in the learning image is discontinuous or an area with fabric wrinkles is detected, and an area other than the area in the learning image that includes the fabric pattern is cut out. A first image cutout unit;
A first feature amount extraction unit that extracts a plurality of feature amounts indicating features of a pattern included in the image cut out by the first image cutout unit;
A representative vector extraction unit that extracts a representative vector for each feature amount obtained by connecting a plurality of feature amounts extracted by the first feature amount extraction unit;
A first feature histogram generation unit that generates a histogram of feature amounts corresponding to the representative vector based on the representative vector extracted by the representative vector extraction unit;
Based on the histogram generated by the first feature histogram generation unit and a label corresponding to the image, an identification parameter that is a weighting factor for the histogram is calculated, and the calculated identification parameter and a representative corresponding to the identification parameter An image classification apparatus comprising: a learning unit that generates pattern classification data associated with a vector. A region where the fabric pattern captured in the input image is discontinuous or a region where the fabric wrinkle is detected is detected, and a region other than the region in the image and including the fabric pattern is cut out. A second image cutout unit;
A second feature amount extraction unit that extracts a plurality of feature amounts indicating features of a pattern included in the image cut out by the second image cutout unit;
A histogram is generated for each of a plurality of feature amounts extracted by the second feature amount extraction unit based on a representative vector included in the pattern classification data generated by the image classification device according to claim 1 or 2. A second feature histogram generator;
An identification unit that identifies a fabric pattern of the input image based on an identification parameter included in pattern classification data stored in the image classification device and a histogram generated by the second feature histogram generation unit And an image identification device.   A program for causing a computer to function as the image classification device according to claim 1.   A program for causing a computer to function as the image identification device according to claim 3.